ocrodeg

The ocrodeg package is a small Python library implementing document image degradation for data augmentation for handwriting recognition and OCR applications.

The following illustrates the kinds of degradations available from ocrodeg.

%pylab inline

Populating the interactive namespace from numpy and matplotlib

rc("image", cmap="gray", interpolation="bicubic")
figsize(10, 10)
import scipy.ndimage as ndi
import ocrodeg

image = imread("testdata/W1P0.png")
imshow(image)

<matplotlib.image.AxesImage at 0x7fabcc7ab390>

PAGE ROTATION

This is just for illustration; for large page rotations, you can just use ndimage.

for i, angle in enumerate([0, 90, 180, 270]):
    subplot(2, 2, i+1)
    imshow(ndi.rotate(image, angle))

RANDOM GEOMETRIC TRANSFORMATIONS

random_transform generates random transformation parameters that work reasonably well for document image degradation. You can override the ranges used by each of these parameters by keyword arguments.

ocrodeg.random_transform()

{'angle': -0.016783842893063807,
 'aniso': 0.805280370671964,
 'scale': 0.9709145529604223,
 'translation': (0.014319657859164045, 0.03676897986267606)}

Here are four samples generated by random transforms.

for i in xrange(4):
    subplot(2, 2, i+1)
    imshow(ocrodeg.transform_image(image, **ocrodeg.random_transform()))

You can use transform_image directly with the different parameters to get a feel for the ranges and effects of these parameters.

for i, angle in enumerate([-2, -1, 0, 1]):
    subplot(2, 2, i+1)
    imshow(ocrodeg.transform_image(image, angle=angle*pi/180))

for i, angle in enumerate([-2, -1, 0, 1]):
    subplot(2, 2, i+1)
    imshow(ocrodeg.transform_image(image, angle=angle*pi/180)[1000:1500, 750:1250])

for i, aniso in enumerate([0.5, 1.0, 1.5, 2.0]):
    subplot(2, 2, i+1)
    imshow(ocrodeg.transform_image(image, aniso=aniso))

for i, aniso in enumerate([0.5, 1.0, 1.5, 2.0]):
    subplot(2, 2, i+1)
    imshow(ocrodeg.transform_image(image, aniso=aniso)[1000:1500, 750:1250])

for i, scale in enumerate([0.5, 0.9, 1.0, 2.0]):
    subplot(2, 2, i+1)
    imshow(ocrodeg.transform_image(image, scale=scale))

for i, scale in enumerate([0.5, 0.9, 1.0, 2.0]):
    subplot(2, 2, i+1)
    h, w = image.shape
    imshow(ocrodeg.transform_image(image, scale=scale)[h//2-200:h//2+200, w//3-200:w//3+200])

RANDOM DISTORTIONS

Pages often also have a small degree of warping. This can be modeled by random distortions. Very small and noisy random distortions also model ink spread, while large 1D random distortions model paper curl.

for i, sigma in enumerate([1.0, 2.0, 5.0, 20.0]):
    subplot(2, 2, i+1)
    noise = ocrodeg.bounded_gaussian_noise(image.shape, sigma, 5.0)
    distorted = ocrodeg.distort_with_noise(image, noise)
    h, w = image.shape
    imshow(distorted[h//2-200:h//2+200, w//3-200:w//3+200])

RULED SURFACE DISTORTIONS

for i, mag in enumerate([5.0, 20.0, 100.0, 200.0]):
    subplot(2, 2, i+1)
    noise = ocrodeg.noise_distort1d(image.shape, magnitude=mag)
    distorted = ocrodeg.distort_with_noise(image, noise)
    h, w = image.shape
    imshow(distorted[:1500])

BLUR, THRESHOLDING, NOISE

There are a range of utilities for modeling imaging artifacts: blurring, noise, inkspread.

patch = image[1900:2156, 1000:1256]
imshow(patch)

<matplotlib.image.AxesImage at 0x7fabc88c7e10>

for i, s in enumerate([0, 1, 2, 4]):
    subplot(2, 2, i+1)
    blurred = ndi.gaussian_filter(patch, s)
    imshow(blurred)

for i, s in enumerate([0, 1, 2, 4]):
    subplot(2, 2, i+1)
    blurred = ndi.gaussian_filter(patch, s)
    thresholded = 1.0*(blurred>0.5)
    imshow(thresholded)

reload(ocrodeg)
for i, s in enumerate([0.0, 1.0, 2.0, 4.0]):
    subplot(2, 2, i+1)
    blurred = ocrodeg.binary_blur(patch, s)
    imshow(blurred)

for i, s in enumerate([0.0, 0.1, 0.2, 0.3]):
    subplot(2, 2, i+1)
    blurred = ocrodeg.binary_blur(patch, 2.0, noise=s)
    imshow(blurred)

MULTISCALE NOISE

reload(ocrodeg)
for i in range(4):
    noisy = ocrodeg.make_multiscale_noise_uniform((512, 512))
    subplot(2, 2, i+1); imshow(noisy, vmin=0, vmax=1)

RANDOM BLOBS

for i, s in enumerate([2, 5, 10, 20]):
    subplot(2, 2, i+1)
    imshow(ocrodeg.random_blobs(patch.shape, 3e-4, s))

reload(ocrodeg)
blotched = ocrodeg.random_blotches(patch, 3e-4, 1e-4)
#blotched = minimum(maximum(patch, ocrodeg.random_blobs(patch.shape, 30, 10)), 1-ocrodeg.random_blobs(patch.shape, 15, 8))
subplot(121); imshow(patch); subplot(122); imshow(blotched)

<matplotlib.image.AxesImage at 0x7fabc8a35490>

FIBROUS NOISE

imshow(ocrodeg.make_fibrous_image((256, 256), 700, 300, 0.01))

<matplotlib.image.AxesImage at 0x7fabc8852450>

FOREGROUND / BACKGROUND SELECTION

subplot(121); imshow(patch); subplot(122); imshow(ocrodeg.printlike_multiscale(patch))

<matplotlib.image.AxesImage at 0x7fabc8676d90>

subplot(121); imshow(patch); subplot(122); imshow(ocrodeg.printlike_fibrous(patch))

<matplotlib.image.AxesImage at 0x7fabc8d1b250>

NVlabs/ocrodeg